Objective unsupervised learning of behavioral dynamics at the frontier of information bottleneck principles, physics informed priors, and maximum mean discrepancy statistical testing

Abstract

When studying model organisms in neuroscience, it is customary to observe their response to various sensory stimuli. A common way to characterize this response is to classify it into a set of discrete behaviors. Those behaviors are defined by specialists after careful observation. Recently, a growing body of literature aims to discover relevant segmentations of the output using unsupervised machine-learning techniques, allowing for objective definitions of a behavior (Berman et al., 2014; Luxem et al., 2020).